Dual-domain prior unfolding network for remote sensing image super-resolution

Super-Resolution (SR) in remote sensing has garnered significant interest due to its ability to improve the spatial resolution and overall quality of images. In recent years, deep-learning-based SR methods have shown promising performance. However, these methods do not incorporate prior knowledge in the design of neural networks and lack interpretability. Deep unfolding methods address these issues by developing networks based on SR models with priors. In this paper, we propose a Dual-Domain Prior Unfolding Network (DDPUNet) for the SR task of remote sensing images. We formulate the SR problem as a general model with priors in both image domain and gradient domain, and then decouple this model as three sub-problems. Each sub-problem is tackled with neural modules, and the unfolding network is developed based on these modules. To learn distinct image features, we use networks with different structures for the learning of priors in the image domain and the gradient domain. The proposed approach outperforms deep-learning-based approaches and other deep unfolding approaches in both objective metrics and image visual quality. Experiments on remote sensing and natural image datasets demonstrate the performance of the proposed method in reconstructing edges and texture details of images.